RPS 605a - Artificial intelligence and MRI radiomics

Purpose:

To achieve high diagnostic precision in multiple sclerosis (MS) and to enable fast clinical decision support, it is necessary to monitor brain injury appearing in white matter (WM) with a reliable and robust method. We investigated the ability of a radiomics models to differentiate between MS patients and healthy subjects (HS) based on clinical T1w and quantitative MRI (qMRI) of WM.

Methods and materials:

A dataset containing T1w and qMRI (R1, R2*, MT, PD maps) of 36 MS patients and 37 HS was used for training. Two external datasets containing T1w images of 359 HS and 15 MS patients were combined for validation. From each MRI scan, 107 radiomics features were extracted and feature selection with bootstrap sampling was performed on the training set. Various binary classification models informed by T1w and both single and mixed qMRI features were trained. Models were assessed using accuracy and “area under the receiver operator characteristic curve” (AUC) metrics. Classifiers based on T1w were tested on the external datasets using resampling, while qMRI models were tested on the training set using a standard cross-validation.

Results:

The 5 most informative and reproducible features were selected for T1w and for quantitative MRI data. On the training (and external) datasets, the best performance was achieved by a support vector classifier using mixed qMRI features.

Conclusion:

Brain WM radiomics features extracted from MRI can be used for automatic status estimation in multiple sclerosis. The best classification performance is achieved with mixed qMRI models. Models are based on simple features, which makes them easily interpretable.

Limitations:

For external validation, available qMRI data is needed.

Ethics committee approval

The study was approved by the local ethic committee (approval B707201213806). Written informed consent was obtained from all participants.

Funding:

Maastricht-Liege Imaging Valley, FRS-FNRS Belgium.

Purpose:

MRI-based tumour response prediction to neoadjuvant systemic therapy (NST) in breast cancer patients is increasingly being studied using radiomics with outcomes that appear to be promising. The aim of this study was to systematically review the current literature and reflect on its quality.

Methods and materials:

A systematic literature search was performed using PubMed and EMBASE databases until May 8th, 2019. Abstracts were read and screened by two reviewers independently. The quality of the radiomics workflow of all eligible studies was assessed using the radiomics quality score (RQS). An overview of the methodologies used in all steps of the radiomics workflow and current results are presented.

Results:

16 studies were selected for inclusion with cohort sizes ranging from 35 to 414 patients. The RQS scores varied from 0% to 41.2%. Methodologies used in the radiomics workflow varied greatly, especially for region of interest (ROI) segmentation, features selection, and model development with heterogeneous outcomes as a result. 7 studies applied univariate analysis and 9 studies applied multivariate analysis. The majority of the studies performed their analysis on the pretreatment dynamic contrast-enhanced T1 weighted (DCE T1W) sequence. Entropy was the best performing individual feature, with AUC values ranging from 0.83 to 0.85. The best performing multivariate prediction model, based on logistic regression analysis, scored an AUC of 0.94 in the validation cohort.

Conclusion:

This systematic review revealed large methodological heterogeneity for each step of the MRI-based radiomics workflow, consequently, the (overall promising) results are difficult to compare. Consensus for standardisation of MRI-based radiomics workflow for tumour response prediction to NST in breast cancer patients is needed to further improve research in this field.

Limitations:

Publication bias and data heterogeneity.

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

In order to extract clinically useful information from medical images, it is of utmost importance that extracted features are reproducible and standardised. This study investigates the robustness of radiomics features, extracted using two commonly used radiomics software, with respect to interobserver manual segmentation variability on dynamic contrast-enhanced breast MRI. In addition, the differences in feature robustness and interobserver segmentation variability between easy-to-segment tumours and more challenging ones, as perceived by the expert, were investigated.

Methods and materials:

129 histologically confirmed breast tumours were segmented manually in three dimensions on the first post-contrast T1-weighted MR exam by four observers: a breast radiologist, a radiologist in training, a PhD candidate with a medical degree, and a medical student. Features were extracted using the RadiomiX and the open-source Pyradiomics softwares. Features with an intraclass correlation coefficient (ICC>0.9) were considered robust. Interobserver variability was evaluated using the volumetric Dice similarity coefficient (DSC).

Results:

The mean DSC for all tumours was 0.81 (range 0.19-0.96). The mean DSC was higher for the easy tumours compared to the challenging tumours (0.83 vs 0.75, respectively, p<0.001). In total, 41.6% (552/1328) and 32.8% (273/833) of all RadiomiX and Pyradiomics features were identified as robust, respectively. Of the 94 easy tumours, 57.5% RadiomiX and 35.7% Pyradiomics features were robust. Of the 35 challenging tumours, only 17.2% RadiomiX and 28.6% Pyradiomics features were robust.

Conclusion:

This study shows conclusively the intuitive notion that more complex, challenging tumours lead to less robust features, and that robust features are not simply interchangeable between radiomics software with respect to interobserver segmentation variability. Ultimately, a list of robust radiomics features which is independent of interobserver segmentation variability in breast MRI was identified for two commonly used software.

Limitations:

n/a

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

To develop and evaluate a MRI derived machine learning radiomics method to preoperatively differentiate gliosarcoma (GSM) from glioblastoma (GBM) and predict survival.

Methods and materials:

A single-centre retrospective review was conducted on 39 GSM patients and 316 GBM patients [training cohort: GSM n=27, GBM n=221; validation cohort: GSM n=12, GBM n=95]. Machine learning radiomics models were built by combining the least absolute shrinkage and selection operator (LASSO), random forest (RF), and support vector machine (SVM) methods.

Results:

The final SVM model provided an AUC of 0.9420 (95% CI: 0.8922–0.9918) with an accuracy of 0.9280 (95% CI: 0.8119-0.9760) in the training cohort and an AUC of 0.9447 (95% CI: 0.9025–0.9870) with an accuracy of 0.8972 (95% CI: 0.8318-0.9533) in the validation cohort. A radiomics score (Rad-score) was calculated which was computed based on Cox regression coefficients of the multivariate Cox regression. All GSM patients were separated into two groups, the low and high risk groups, which are the patients with Risk-score ≤median and Risk-score > median, respectively. The result of a lokrank test indicated that the median survival of the two groups was statistically significant (P = 0.00007).

Conclusion:

This study presents a pre-enhanced MRI image-based machine learning radiomics method that could potentially facilitate the preoperative identification of GSM tumour from GBM and predict the prognosis for GSM.

Limitations:

Our study was a retrospective study conducted in a single institution. Therefore, there might be case selection bias and a shortage of accurate radiology-pathology correlation study to reveal the underlying pathology.

Ethics committee approval

Our Institutional Review Board approved this retrospective study (No. 201709995) and waived informed consents as this study was retrospective.

Funding:

This study is partially supported by China Postdoctoral Science Foundation funded project (2018M632997).

Purpose:

To evaluate radiomics performance in comparison to radiological visual assessment for the discrimination between ≤T2 and ≥T3 rectal cancer tumours on T2-weighted (T2w) MR images.

Methods and materials:

This multi-institutional study included two datasets of consecutive patients with rectal cancer who underwent total mesorectal excision as primary curative treatment, without neoadjuvant therapy (n=23-institution A; n=20-institution B). Two patient groups were formed based on pathological stage, <=T2 (n=22) and >=T3 (n=21). Patients underwent staging MRI before surgery comprising high-resolution T2w imaging. T2w images were blindly reviewed by 2 radiologists and primary tumours were classified as <=T2 or >=T3. A volume of interest around the whole primary tumour was drawn by the 2 radiologists and the corresponding datasets were analysed using PyRadiomics. Feature stability was tested using a correlation coefficient cutoff of 0.75. Feature reduction was performed, excluding highly correlated features. A general linear model (w/LASSO) was then run to select the best performance feature. Repeated cross-validation was performed. Interobserver agreement was estimated for T2w visual assessment by the 2 radiologists. A DeLong test was employed to assess if the differences between auROC were statistically significant.

Results:

'Spherical disproportion' was the best discriminator with an accuracy upon cross-validation of 72.3% +- 12.6%. The accuracy of the T2w visual analysis was 63% and 77% for the two reader with an intraclass correlation coefficient of 0.71 [95CI:0.46-0.84]. Comparison between radiomics analysis and visual analysis revealed no statistically significant differences (p=0.95 and p=0.14).

Conclusion:

The performance of radiomics analysis based on T2w is similar to that of visual assessment by radiologists.

Limitations:

An additional study to include radiomic features of the surroundings of the tumour might lead to better staging discrimination.

Ethics committee approval

Ethics committee approval obtained.

Funding:

No funding was received for this work.

Purpose:

To identify potential imaging biomarkers of SDHx mutation in paragangliomas and pheochromocytomas on MRI using radiomics.

Methods and materials:

Mice with a subcutaneous tumour carrying a homozygous knockout of the SDHB gene (SDHB^-/-) or the wild type (WT) counterpart (SDHB^lox/lox) were studied on MRI, respectively 22 SDHB^-/- and 16 WT tumours. Analysis of the True Fisp-weighted images using a radiomics workflow allowed for the extraction of 105 features. Feature reduction based on reproducibility and statistical analysis evaluated the performance of each feature to detect the SDHB mutation. The candidate features were then tested in patients with paragangliomas or pheochromocytomas: 9 with SDHx mutations, 10 with no mutation, and 13 with a mutation other than SDHx.

Results:

In the 38 mice studied by MRI, 17 features were significantly different between the SDHB^-/- tumours and WT tumours, using a Benjamini and Hochberg correction with a False Discovery Rate of 0.1.

When the 17 features were tested in the 32 patients, only 2 remained significantly associated to the presence of the SDHx mutation in humans: 1) the glcm_JointEnergy (0.009 vs 0.004; p=0.03 between SDHx mutations and no mutation and p=0.003 between SDHx mutations and any mutation other than SDHx or no mutation) and 2) the glcm_JointEntropy (7.9 vs 8.6 ; p=0.01 between SDHx mutations and any mutation other than SDHx or no mutation).

Conclusion:

Using radiomics and a data-driven and big-data method on MR images, a texture feature reflecting tumour heterogeneity was shown to be predictive of SDHx-mutation status both in a mouse model and in patients.

Limitations:

Few SDHB patients. The Fiesta sequence is not systematic in the MRI protocol of paragangliomas and pheochromocytomas. A monocentric study.

Ethics committee approval

Ethical and regulatory authorisations were obtained for both the pre-clinical and clinical studies.

Funding:

No funding was received for this work.

Purpose:

To characterise mesorectal lymph nodes (LNs) extracted from rectal cancer patients employing radiomics analysis on gradient-echo MRI images.

Methods and materials:

29 benign and 35 malignant LNs retrieved from 11N+ rectal cancer patient specimens were examined in a 16.4T scanner. A T2* multi-gradient-echo sequence was acquired (50 TEs starting at 1.6ms and 1.4ms interval). Datasets were denoised while a radiologist segmented the central slice of each node and the third echo was considered for radiomics analysis. 757 radiomics features were extracted from the original, Laplacian of Gaussian, and wavelet images. Feature extraction was performed using pyradiomics. Stratified partition was used to divide data into training (80%-52 LNs) and test sets (20%-12 LNs). Feature reduction was performed by removing near-zero and zero variance, as well as highly correlated features. Univariate analysis was performed while the Bonferroni test was employed for multiple comparisons correction. Significant features were used to create a logistic regression model, which was optimised using 100-times repeated 5-fold cross-validation. The fine-tuned model was applied to the test set and the area under the receiver operating characteristic curve (auROC) was calculated.

Results:

The 5 features selected from the training set had auROCs between 0.79 and 0.82. After repeated cross-validation, the logistic regression model applied on the training set provided a mean auROC of 0.81, with a standard deviation of 0.12. The optimised logistic regression model applied to the test set provided an auROC of 0.89.

Conclusion:

Radiomics analysis based on gradient-echo MRI has excellent performance for the distinction between benign and malignant mesorectal LNs extracted from rectal cancer patients.

Limitations:

The low number of patients might be a limitation of this work.

Ethics committee approval

Ethics committee approval obtained.

Funding:

No funding was received for this work.

Purpose:

Although several studies have investigated the reproducibility of radiomics data derived from CT and PET/CT, data on the reproducibility of MR-based radiomics are scarce. This study aims to assess the reproducibility of radiomic features derived from pelvic MRI data and study the effects of variations between readers, segmentation methodology, and feature extraction software packages.

Methods and materials:

25 pelvic MRIs (T2W-MRI of anal cancer) were retrospectively analysed and segmented by two readers to include the: [1] whole-tumour volume and [2] largest single axial tumour-slice. Pixel intensities were normalised to mean=300/SD=100 and images were resampled isotropically (2x2x2mm³). Radiomic features were extracted using 2 open-source packages (PyRadiomics-v2.2.0, CaPTk-v1.7.3), using comparable settings without image filtration. Segmented pixel intensities were quantised using a fixed bin width of 5. Only features defined in both packages were extracted (first-order, shape, GLCM, GLRLM, GLSZM, and NGTDM features, 51 total). For each feature, the intraclass correlation coefficient (ICC) was calculated between the [1] two readers, [2] two segmentation methods (whole-volume vs. single-slice), and [3] two software packages. When comparing segmentation methods, shape features were excluded from the analysis.

Results:

Inter-reader reproducibility was moderate (20/51 features;0.5<ICC<=0.75) to good (15/51 features;0.75<ICC<=0.9). Between segmentation methods, the majority of features (in particular GLRLM, GLSZM, and NGTDM) showed poor reproducibility (31/45 features; ICC<0.5), though most first-order features showed good (7/15 features; 0.75<ICC<=0.9) to excellent (2/15 features; ICC>0.9) reproducibility. Between software packages, the majority of first-order, shape, GLCM, and GLRLM features showed excellent reproducibility (23/30 features; ICC>0.9). The remaining higher-order features (GLSZM and NGTDM) were all poorly reproducible (21/21 features; ICC<0.5).

Conclusion:

Variations in software and segmentation methodology negatively affected measurement reproducibility in MRI-based radiomics, in particular for higher-order features. Inter-reader reproducibility was moderate-good.

Limitations:

n/a

Ethics committee approval

n/a

Funding:

No funding was received for this work.

Purpose:

Deep myometrial invasion (≥ 50% wall thickness, DMI) is the most important morphological prognostic factor for endometrial cancer (EC). While DMI can be assessed on MRI, it often proves challenging, reader-experience dependent, and with interobserver variability. Radiomics allow quantification of tumour heterogeneity and has been successfully paired to machine learning (ML). We aimed to detect DMI in EC patients using radiomics-powered ML.

Methods and materials:

One operator manually segmented lesion volumes of interest (VOIs) to extract radiomic features from T2-weighted images. Two additional readers performed lesion segmentation on 30 random patients to test feature stability. Patients were randomly split into model development and validation sets. Multistep feature reduction was performed on the first set. Only stable features (interobserver correlation coefficient ≥ 0.75) were employed. A variance threshold ≤ 0.01 was applied. Highly intercorrelated features (≥ 0.80) were also discarded. Finally, a random forest wrapper was used to select the most significant features. These were employed to train and test via 10-fold cross-validation an ensemble algorithm, a bagged J48 decision tree, whose performance was also assessed on the validation set.

Results:

Retrospectively, 54 patients were identified (DMI histopathologically was proven in 17). 1,132 features were extracted and 144 were stable. Then, 21 were discarded due to low variance. After the removal of the highly intercorrelated ones, the set was reduced to 19 features. The random forest wrapper finally identified the 3 most useful ones. In the cross-validation testing, an 86% accuracy was obtained with an AUC of 0.92. In the validation set, these were respectively 91% and 0.94.

Conclusion:

Radiomics-powered ML appears as a promising tool to accurately detect DMI in EC patients on MRI.

Limitations:

No external validation and a small population.

Ethics committee approval

IRB approved and consent waived.

Funding:

No funding was received for this work.

Purpose:

Pituitary macroadenomas are the most common tumours of the pituitary gland. As far as they are usually soft masses, some of them may present with a harder consistency, making it difficult to remove them by using their standard surgical technique, transsphenoidal adenomectomy. The objective of the following study was to determine the accuracy of machine learning texture analysis in assessing the consistency of pituitary macroadenomas in patients subdued to endoscopic endonasal surgery.

Methods and materials:

A total of 89 patients (68 soft, 11 rubbery, and 10 fibrous adenomas) who underwent an endoscopic endonasal procedure for pituitary adenoma removal were retrospectively included. Two readers independently segmented the lesions on T2-weighted MR images for the extraction and stability testing of radiomic features. Oversampling of the minority classes was used to balance the data. A 75/25% training-testing split was used for model validation. During feature selection on the training data, unstable, low-variance, and highly-intercorrelated parameters were removed. Subsequently, the most informative features were identified with a random forest-based wrapper and a random forest was employed for consistency prediction in the validation set.

Results:

Of the 1,118 features extracted, 599 were stable with an ICC >= 0.90. Only one of these showed a variance <= 0.01, while 526 were highly intercorrelated (>= 0.80). Among the remaining 72 features, the wrapper selected 10 as most useful. The finalised random forest model achieved an overall accuracy of 84.31% with an AUC of 0.97.

Conclusion:

Radiomics and machine learning showed high accuracy in the prediction of pituitary adenoma consistency on pre-operative T2-weighted MR images, promising to be a useful tool in the surgical approach choice.

Limitations:

A single institution study.

Ethics committee approval

IRB approved. Written informed consent was waived.

Funding:

No funding was received for this work.

Purpose:

Prognostic markers calculated from diagnostic imaging, also called radiomics, have shown promising results in predicting treatment outcomes of different cancer types. This study assessed MR-based radiomics to predict locoregional failure (LRF) after chemoradiation treatment in oropharyngeal squamous cell carcinoma (OPSCC) patients. The added value of radiomic features to traditional clinical outcome predictors was evaluated.

Methods and materials:

177 patients were included in the study, 76 (43%) HPV positive tumours and 77 (44%) negative HPV tumours. Pretreatment MRI, clinical variables, and outcome data were retrospectively collected from OPSCC patients. In addition, 824 radiomic features were extracted within the manually delineated primary tumour region. Three machine learning models based on either clinical variables (clinical model), radiomic features (radiomic model), and both combined (combined model) were created. The latter two were constructed using wrapper feature selection in combination with random forest, while the clinical model was based on a support vector machine. Prediction quality was assessed using area under the curve (AUC). Different models were compared using the McNeil test. All analyses were performed for the total population and stratified for the tumour human papillomavirus (HPV) status.

Results:

Within the total patient cohort, the radiomic model showed the best predictive performance with an AUC of 0.65 compared to the clinical model (AUC: 0.41, p=0.04) and combined model (AUC: 0.63, p=0.07). The same trend was seen in HPV subgroups, however, without reaching statistical significance.

Conclusion:

MR-based radiomics is able to predict LRF after chemoradiation in OPSCC patients and performed better compared to clinical variables.

Limitations:

All patients were selected from a single-centre while applying radiomics in multicentre cohorts is preferred. Another limitation is the single reader design of the delineations.

Ethics committee approval

IRBd18047.

Funding:

Verwelius Foundation.